A transcript is the full record of an AI agent's run, including its messages, the tool calls it made and the outputs it received. It shows what the agent did step by step, not only the final result.
It matters because the final answer alone can hide how a result was reached. Reading transcripts shows why an agent failed, whether it succeeded for the right reasons and whether the grader scored it fairly. They are also the raw material for measures such as the number of steps or tool calls.
Transcripts can be long, and reading many of them by hand takes time, so teams often sample or use automated analysis. They can also contain sensitive data, such as customer information or credentials, which should be handled with care when stored or shared. A transcript records behavior and does not by itself prove what the model was reasoning.
To make use of them, save them for every trial, review a sample of both successes and failures and protect or remove any sensitive content before sharing.